Evo 2 lands in Nature

Evo 2 matters because it pushes “language model” thinking beyond text and into programmable biology, with scale numbers that look much closer to frontier LLMs than niche bioinformatics models.

• –The model’s 1M-context genomic window is the real technical unlock, since long-range dependencies are essential for understanding regulation and genome structure
• –Arc is treating Evo 2 like infrastructure, not just a paper: code, weights, data, GitHub access, Hugging Face distribution, and BioNeMo integration all lower the barrier for researchers to build on top
• –The strongest developer takeaway is that genomics is starting to inherit the same stack patterns as mainstream AI: foundation models, open weights, fine-tuning, tooling layers, and interpretability work
• –Safety is part of the story too, with Arc explicitly excluding eukaryotic viruses from training and reporting red-team results around pathogenic sequence generation